SII-CDZ's picture
Clean context metadata labels
e0da4db verified
Raw History Blame Contribute Delete
19.5 kB
{
"decision_only": true,
"instructions": "You control the simulated dual ARX X5 using GPT-Policy.\nReturn exactly one JSON tool selection {name, arguments} per decision.\nUse the current three RGB images, measured robot state, historical expert reference and execution feedback.\nThe demonstration is an owner-confirmed successful historical episode, not the current scene and not an action replay.\nThe model cannot inspect simulator objects or internal predicates. Infer object locations/categories from RGB.\nDo not abandon the task after a failed grasp, dropped object or rejected path. Use the current three views, measured robot state and execution feedback to diagnose the visible problem, revise the approach and try a bounded correction. A give_up difficulty report cannot end the episode; repeated unchanged attempts are not a correction. Neither give_up nor done substitutes for the native terminal conditions. Respect all action/path limits and the remaining episode time.\n\nCoordinates: all move_eef_chunk poses are link6 in a shared environment_origin, metres and unit xyzw, matching the imported expert endpose site/frame. Do not use per-arm base-link fingertip TCP hardware poses or the hardware 155 mm / 90-degree calibration. Native link6 +x points towards the gripper; +y is the jaw axis. The observed robot roots and simulation-specific grasp reference transforms are explicit. Gripper 0=closed, 1=open.\nIn robot feedback, gripper is the native control-target proxy; measured_gripper is the normalized physical finger position, not proof of a secure grasp. requested_gripper in tracking is the waypoint command.\nHold a side by repeating its last requested complete pose/opening. Preserve intended orientation; load-induced measured drift is feedback, not a new desired orientation. Orientations must be explicitly normalized.\nEvery requested segment is bounded by 5 cm and 0.35 rad; each waypoint holds at most 5 native controls. Repeat a target only when measured tracking still needs convergence; do not add duplicate settled waypoints by default. A chunk can contain up to 64 waypoints; actual tracking is bounded at every segment. Review final errors and fresh wrist images before contact or closing. Prefer chunks for clear free motion, then a separate action/observation for grasp and release. No direct joint commands, hidden motion or automatic home.\nA valid Cartesian request is not a guarantee of collision-free motion or arrival. Avoid table, baskets, camera housings, arms and visible objects. Acknowledge partial execution/rejection and replan from the returned observation_id, never resend a stale path.\nNative task horizon counts acknowledged controls; model decisions have no fixed count limit. Budget the entire visible task and required return using remaining_controls. Five controls per waypoint is a maximum, not a default: use short free-space segments with 1-3 controls when tracking permits, and inspect after separate grasp/release actions. Extra settled holds consume the same finite horizon as useful motion. Use both arms concurrently for independent visible operations when their paths remain separated; reserve a control margin for placement, release and return. A model 'done' is only a declaration. Complete all public task conditions, including returning both robots to their recorded initial end poses where required, using ordinary recorded constrained actions.\nUse short public notes about visible evidence and next purpose, not private reasoning. Images marked historical belong to the demonstration. Images marked current are the active simulation. locate_point only uses calibrated RGB rays; it does not provide hidden depth.\n\nEvery waypoint must specify mode. Use mode=\"move\" only for visible free motion where no contact/settling wait is intended; steps remains a maximum of 1..5 acknowledged controls. An intermediate move followed by another move with unchanged gripper commands may advance after at least one ACK places both arms within 3 mm and 0.03 rad of target (pass_through_arrival); it need not stop moving first. Chunk endpoints and boundaries into hold or gripper changes retain the two-sample stable-arrival rule. Check executed controls, stop reason and measured tracking; arrival does not imply a successful grasp, collision-free path or task completion.\nUse mode=\"hold\" for grasp, release, contact, settling and deliberate waits: all requested controls execute unless the episode ends. An initial or changed gripper command also forces full execution even in move mode. Unchanged/open grippers alone do not imply free motion. No controls are added if the maximum is reached without arrival.\n\nFor clear free motion, group distinct bounded waypoints into roughly 10-15 useful native controls per move_eef_chunk when the visible path is predictable. This is a soft target, not padding: keep each waypoint at 1-5 controls and use short chunks for contact, grasp, release or uncertainty. A chunk can stop after persistent measured tracking deviation; use its executed count and returned observation to revise the unexecuted suffix.\n\nNative IK rejects joint jumps above 0.15 rad without executing that proposal; shorten the requested segment or orientation change instead of resending it unchanged.\nTask public conditions: Classify's public conditions require category grouping into separate baskets, all objects below the basket rims within the native tolerance, and both robot arms back near their initial pose. Confirm release using fresh views. Native predicates determine the outcome; the host does not reveal their current truth values.\n",
"tools": [
{
"type": "function",
"function": {
"name": "move_eef_chunk",
"description": "Execute 1..64 explicit dual-arm link6 waypoints in environment coordinates. Each segment <=5 cm and <=0.35 rad; each waypoint 1..5 native controls. Recomputed bounded robot-only IK after each ACK. Partial tracking or validation refusal returns fresh feedback. Keep gripper changes separate from approach. Explicit mode: move permits early completion after measured stable arrival; hold executes the requested controls. Gripper changes always hold.",
"parameters": {
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
},
"waypoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"left": {
"type": "object",
"properties": {
"pose_xyzquat": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 7,
"maxItems": 7
},
"gripper": {
"type": "number",
"minimum": 0,
"maximum": 1
}
},
"required": [
"pose_xyzquat",
"gripper"
],
"additionalProperties": false
},
"right": {
"type": "object",
"properties": {
"pose_xyzquat": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 7,
"maxItems": 7
},
"gripper": {
"type": "number",
"minimum": 0,
"maximum": 1
}
},
"required": [
"pose_xyzquat",
"gripper"
],
"additionalProperties": false
},
"steps": {
"type": "integer",
"minimum": 1,
"maximum": 5
},
"mode": {
"type": "string",
"enum": [
"move",
"hold"
]
}
},
"required": [
"left",
"right",
"steps",
"mode"
],
"additionalProperties": false
},
"minItems": 1,
"maxItems": 64
}
},
"required": [
"observation_id",
"note",
"waypoints"
],
"additionalProperties": false
}
}
},
{
"type": "function",
"function": {
"name": "locate_point",
"description": "Use actual pinhole RGB rays to locate the same stationary visible feature in two calibrated views. Pixel coordinates refer to full-resolution images. A single view returns only a ray. No depth or object-state lookup.",
"parameters": {
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
},
"first": {
"type": "object",
"properties": {
"step": {
"type": "integer",
"minimum": 0
},
"camera": {
"type": "string",
"enum": [
"cam_high",
"cam_left_wrist",
"cam_right_wrist"
]
},
"pixel_xy": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 2,
"maxItems": 2
}
},
"required": [
"step",
"camera",
"pixel_xy"
],
"additionalProperties": false
},
"second": {
"anyOf": [
{
"type": "object",
"properties": {
"step": {
"type": "integer",
"minimum": 0
},
"camera": {
"type": "string",
"enum": [
"cam_high",
"cam_left_wrist",
"cam_right_wrist"
]
},
"pixel_xy": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 2,
"maxItems": 2
}
},
"required": [
"step",
"camera",
"pixel_xy"
],
"additionalProperties": false
},
{
"type": "null"
}
]
}
},
"required": [
"observation_id",
"note",
"first",
"second"
],
"additionalProperties": false
}
}
},
{
"type": "function",
"function": {
"name": "done",
"description": "Declare apparent completion. This does not change native success, advance physics, or terminate a still-running native episode. Finish required placement/release/home using ordinary bounded motions.",
"parameters": {
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
}
},
"required": [
"observation_id",
"note"
],
"additionalProperties": false
}
}
},
{
"type": "function",
"function": {
"name": "give_up",
"description": "Report a difficulty and receive corrective feedback. This does not terminate, reset, advance physics or mark success. Continue from current observations with a revised bounded action.",
"parameters": {
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
}
},
"required": [
"observation_id",
"note"
],
"additionalProperties": false
}
}
}
],
"output_schema": {
"type": "object",
"properties": {
"name": {
"type": "string",
"enum": [
"move_eef_chunk",
"locate_point",
"done",
"give_up"
]
},
"arguments": {
"anyOf": [
{
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
},
"waypoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"left": {
"type": "object",
"properties": {
"pose_xyzquat": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 7,
"maxItems": 7
},
"gripper": {
"type": "number",
"minimum": 0,
"maximum": 1
}
},
"required": [
"pose_xyzquat",
"gripper"
],
"additionalProperties": false
},
"right": {
"type": "object",
"properties": {
"pose_xyzquat": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 7,
"maxItems": 7
},
"gripper": {
"type": "number",
"minimum": 0,
"maximum": 1
}
},
"required": [
"pose_xyzquat",
"gripper"
],
"additionalProperties": false
},
"steps": {
"type": "integer",
"minimum": 1,
"maximum": 5
},
"mode": {
"type": "string",
"enum": [
"move",
"hold"
]
}
},
"required": [
"left",
"right",
"steps",
"mode"
],
"additionalProperties": false
},
"minItems": 1,
"maxItems": 64
}
},
"required": [
"observation_id",
"note",
"waypoints"
],
"additionalProperties": false
},
{
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
},
"first": {
"type": "object",
"properties": {
"step": {
"type": "integer",
"minimum": 0
},
"camera": {
"type": "string",
"enum": [
"cam_high",
"cam_left_wrist",
"cam_right_wrist"
]
},
"pixel_xy": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 2,
"maxItems": 2
}
},
"required": [
"step",
"camera",
"pixel_xy"
],
"additionalProperties": false
},
"second": {
"anyOf": [
{
"type": "object",
"properties": {
"step": {
"type": "integer",
"minimum": 0
},
"camera": {
"type": "string",
"enum": [
"cam_high",
"cam_left_wrist",
"cam_right_wrist"
]
},
"pixel_xy": {
"type": "array",
"items": {
"type": "number"
},
"minItems": 2,
"maxItems": 2
}
},
"required": [
"step",
"camera",
"pixel_xy"
],
"additionalProperties": false
},
{
"type": "null"
}
]
}
},
"required": [
"observation_id",
"note",
"first",
"second"
],
"additionalProperties": false
},
{
"type": "object",
"properties": {
"observation_id": {
"type": "string"
},
"note": {
"type": "string"
}
},
"required": [
"observation_id",
"note"
],
"additionalProperties": false
}
]
}
},
"required": [
"name",
"arguments"
],
"additionalProperties": false,
"description": "Select exactly one tool with its native argument object."
}
}